Evolutionary optimization of energy systems using population graphing and neural networks

Abstract

This paper examines the simultaneous use of graph based evolutionary algorithms (GBEAs) and a real-time estimate of the final fitness for evolutionary optimization of systems modeled using computational fluid dynamics (CFDs). GBEAs are used to control the rate at which information travels, enabling the diversity of the population to be tuned to match the solution space. During each fitness evaluation, the CFD solver iteratively solves the fluid flow and heat transfer characteristics of the proposed design. In this paper, an artificial neural network is used to develop a real-time estimate of the final fitness and error bounds at each iteration of the solver. Using these estimates, the evolutionary algorithm can determine when the fitness of the design is known with sufficient accuracy for the evolutionary process. This significantly reduces the overall compute time. These techniques are demonstrated by optimizing the spatial temperature profile of the cooking surface of a biomass cookstove. In this cookstove, hot gases from biomass combustion flow under the cooking surface. Within this flow area, a set of baffles direct the flow of hot gases and establish the spatial temperature profile of the stove's cooking surface. The location and size of a series of baffles within the hot gas flow area are determined by the optimization routine. In this design problem, it is found that the two techniques are compatible; both the number of fitness evaluations and the time required for each CFD fitness evaluation are reduced while utilizing GBEAs to preserve the diversity of the population.